Research map / Probabilistic and scientific modeling / Differential equations and PDEs
Neural Physical Simulation with Multi-Resolution Hash Grid Encoding
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a mesh-free neural PDE solver that encodes the solution with a multi-resolution hash grid, using finite-difference derivatives with ghost nodes at boundaries and range-analysis KD-tree sampling for dynamic geometry.
Problem
Physical simulation with neural implicit representations is too compact for local details and converges slowly, grid/mesh methods are memory-heavy and not adaptive, and boundary sampling is inefficient for dynamic geometries.
Approach
Replace the compact MLP or Siren encoder with a multi-resolution hash grid encoder, compute high-order derivatives by finite differences (with ghost-node stencils for boundary pressure gradients) instead of locality-limited analytic gradients, and sample dynamic implicit boundaries via range analysis over a KD-tree.
Key results
- On the multi-scale Taylor-Green flow, the method reaches MSE 8.84e-5 versus 2.02e-4 for Siren, while using about 50% less time (5.6 h vs 10.8 h).
- On the 1D advection equation, the method matches Siren accuracy (MSE 4.2e-7 vs 4.6e-7) with about 40% less runtime (2.30 h vs 5.33 h).
- On the 2D elastic collision example, the method's MSE is comparable to Siren's at 0.3 to 0.7 s and its runtime is 10.7 min versus 18.5 min for Siren over 10 steps.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use multi-resolution hash grid encoding instead of Siren implicit representation | Neural physical simulation · 1D advection | MSE: 4.6e-7 -> 4.2e-7 | No clear effect |
| Use multi-resolution hash grid encoding instead of Siren implicit representation | Neural physical simulation · 1D advection | runtime: 5.33 h -> 2.30 h | Helped |
| Use multi-resolution hash grid encoding instead of Siren implicit representation | Neural physical simulation · 2D collision | runtime: 18.5 m -> 10.7 m | Helped |
| Use multi-resolution hash grid encoding instead of Siren implicit representation | Neural physical simulation · 2D collision | MSE at 0.3 s: 8.2e-4 -> 1.0e-3 | Hurt |
| Use multi-resolution hash grid encoding instead of Siren implicit representation | Neural physical simulation · 2D collision | MSE at 0.5 s: 9.3e-4 -> 1.1e-3 | Hurt |
| Use multi-resolution hash grid encoding instead of Siren implicit representation | Neural physical simulation · 2D collision | MSE at 0.7 s: 2.7e-3 -> 2.2e-3 | Helped |
| Use multi-resolution hash grid encoding instead of Siren implicit representation | Neural physical simulation · multi-scale Taylor-Green flow | MSE: 2.02e-4 -> 8.84e-5 | Helped |
| Use multi-resolution hash grid encoding instead of Siren implicit representation | Neural physical simulation · multi-scale Taylor-Green flow | runtime: 10.8 h -> 5.6 h | Helped |
Surprising findings
- The naive grid baseline at the same finest resolution failed to reach comparable accuracy, so the hash encoding's efficiency depends on the multi-resolution design.
- The method shows energy dissipation in vortex street simulation because it gets trapped in local optima for boundary coefficients, which harms long-term accuracy.
Limitations
- Energy dissipation in long-term vortex street simulation due to local optima in boundary condition coefficients.
- Linear interpolation in the hash grid gives zero second derivatives, so the finite-difference approach is only approximate there.
- No training data is needed, but each simulation requires per-scene optimization, so generalization across scenes is not shown.
Experimental setup
Benchmarks: 1D advection equation 2D elastic tension test 2D elastic collision Taylor-Green vortex multi-scale Taylor-Green flow 2D cylinder wall flow
Metrics: mean square error runtime
Baselines: Siren Traditional grid method PINN PINN with temporal sub-domains physics-informed DeepONet Separable PINN FEM MPM FLUENT
Builds on: Instant-NGP multi-resolution hash encoding Siren PINN neural signed distance function Stam stable fluids
Closest papers by meaning
- Compute-optimal data scaling for neural surrogates via multi-fidelity training (NeurIPS 2026 · similarity 0.86)
- Hermite-NGP: Gradient-Augmented Hash Encoding for Learning PDEs (ICML 2026 · similarity 0.86)
- Transolver++: An Accurate Neural Solver for PDEs on Million-Scale Geometries (ICML 2025 · similarity 0.85)
- Text2PDE: Latent Diffusion Models for Accessible Physics Simulation (ICLR 2025 · similarity 0.84)
- Breakeven complexity: A new perspective on neural partial differential equation solvers (NeurIPS 2026 · similarity 0.84)
- AROMA: Preserving Spatial Structure for Latent PDE Modeling with Local Neural Fields (NeurIPS 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.